dsh-proactive

Most multi-model schedulers are inherently reactive: they fire when a signal arrives and idle when none does, and when things go wrong they just wait for the crash. dsh-proactive layers a proactive-intelligence tier on top of the DeepSeek Harness scheduling loop so the system keeps working even in silence. With no pending task it inspects its own runtime state, spots bottlenecks, and spawns improvement goals; in unfamiliar territory it launches exploratory runs; it forecasts upcoming load to pre-reserve capacity; and when anomalies surface it trips circuit breakers, throttles, and degrades gracefully instead of letting failures cascade.

Three pillars carry the engine. First, a dual-mind core pairs a Scientist (Bayesian optimal experiment design that prices the value of learning itself) with a Theorist (hierarchical Bayesian inference plus Minimum Description Length that compresses observations into reusable laws). Second, a cognitive-energy symbiotic economy turns knowledge, beliefs, and agent contributions into tradeable, auditable, self-correcting assets so underperforming strategies are naturally phased out. Third, a ten-one-step Autonomy Loop orchestrates meta-cognitive observation, world-model prediction, goal generation, policy evolution, and memory distillation on every heartbeat, tempered by Ebbinghaus-style forgetting curves and confidence decay so the system gets sharper with use without drifting into over-fitting.

It is aimed at teams and solo developers already running multi-model workflows on DeepSeek Harness who want the scheduler to drive itself rather than sit and wait. The plugin ships ready to use with zero manual key setup and a full roster of domestic LLM providers out of the box. Through its host-fusion layer it also extends observability and safety governance across the entire host tool pipeline: a global kill switch can freeze every tool call, governance fails closed while observation fails open, and no extra dependencies are introduced.